Method and system for ultra-short-term photovoltaic power prediction based on image-derived graph structure
By using an image-derived graph structure-based approach and fusing features through superpixel segmentation and multi-head attention mechanisms, the error problem of traditional photovoltaic power generation forecasting in complex weather environments is solved, achieving high-precision photovoltaic power forecasting and supporting the stable operation of the power grid.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HOHAI UNIV
- Filing Date
- 2025-04-22
- Publication Date
- 2026-06-16
AI Technical Summary
Traditional photovoltaic power generation forecasting methods struggle to capture the subtle effects of real-time weather changes in complex and variable weather environments, leading to significant forecasting errors that impact the safe operation of the power grid and dispatch strategies.
An ultra-short-term photovoltaic power prediction method based on image-derived graph structure is adopted. Image features are extracted through superpixel segmentation and ResNet neural network to construct dynamic graph structure. Multi-modal feature fusion is performed using multi-head attention mechanism, and the future photovoltaic power value is output by combining time series prediction model.
It improves the accuracy and stability of photovoltaic power forecasting, and can maintain high forecasting performance under complex and changeable weather conditions, supporting the safe and stable operation of the power grid.
Smart Images
Figure CN120411707B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of new energy power generation technology, specifically to an ultra-short-term photovoltaic power prediction method and system based on image-derived graph structure. Background Technology
[0002] With the rapid development of new energy technologies, photovoltaic power generation has become an indispensable and important component of the modern clean energy system. Photovoltaic power generation utilizes solar energy resources and has significant advantages such as being clean, renewable, and pollution-free. However, due to the influence of external environmental factors such as sunlight conditions, temperature, humidity, and meteorological factors, the output of photovoltaic power generation exhibits significant fluctuations and uncertainties.
[0003] This volatility not only affects the output stability of photovoltaic power generation systems but also poses a significant challenge to the safe operation and dispatch strategies of the power grid. Especially with the increasing proportion of photovoltaic power generation, accurate prediction of power output becomes increasingly important. Traditional forecasting methods based on historical data and statistical approaches often struggle to capture the subtle impacts of real-time weather changes on photovoltaic output power in complex and variable weather environments, leading to significant prediction errors.
[0004] In recent years, with the development of big data and artificial intelligence technologies, prediction methods combining images, time-series data, and dynamic environmental information have gradually emerged. In particular, the fusion of multi-source information such as satellite cloud images and ground meteorological monitoring data has provided new ideas for achieving higher-precision ultra-short-term photovoltaic power prediction. How to make full use of multi-source heterogeneous data to establish flexible and robust prediction models has become a key technical challenge for improving prediction accuracy and promoting the large-scale application of photovoltaic power generation. Summary of the Invention
[0005] Purpose of the invention: This invention proposes an ultra-short-term photovoltaic power prediction method and system based on image-derived graph structure, which improves the accuracy of ultra-short-term photovoltaic power prediction.
[0006] Technical Solution: To achieve the above-mentioned objectives, the present invention proposes an ultra-short-term photovoltaic power prediction method based on image-derived graph structure, comprising the following steps:
[0007] The ground cloud image sequence is obtained, and the image features are segmented using the superpixel segmentation method to obtain superpixel blocks. The image spatial features are then extracted from the superpixel blocks using the ResNet neural network.
[0008] Using image spatial features and historical photovoltaic power time series features as nodes, a dynamic graph structure is constructed by adjusting edge weights based on the similarity of node features using a dynamic adjacency matrix.
[0009] Based on the dynamic graph structure, a graph sampling aggregation network is used to extract local neighborhood features and global coupled features in a hierarchical manner. Through mean aggregation and residual connection, dynamic graph structure features are obtained.
[0010] Based on photovoltaic power values, and using time-series features, image spatial features, and dynamic graph structure features as inputs, a multi-head attention mechanism is used to fuse multi-modal features to obtain a fused feature vector.
[0011] The fused feature vectors are used to output photovoltaic power values for a specified time interval within the next 5-30 minutes through a time-series prediction model.
[0012] Furthermore, image feature segmentation is performed using superpixel segmentation methods to obtain superpixel blocks, including:
[0013] Convert the RGB image to the CIELAB color space, and calculate a 5-dimensional feature vector [l,a,b,x,y] to represent each pixel in the image, where (l,a,b) represents the color and (x,y) represents the spatial location vector. The similarity between two pixels is measured by the distance between the 5-dimensional vectors.
[0014] The initial seed point position is adjusted based on gradient optimization, and superpixel blocks are generated by iterative clustering.
[0015] Furthermore, the distance metric formula for clustering is:
[0016]
[0017] Where, d c For color distance, d s S is the spatial distance, C is the balance factor, S is the superpixel side length, and D is the distance between cluster centers.
[0018] Furthermore, using image spatial features and historical photovoltaic power time-series features as nodes, a dynamic graph structure is constructed by adjusting edge weights based on the similarity of node features using a dynamic adjacency matrix, including:
[0019] The feature vector of each superpixel block extracted by the ResNet neural network is used as a node representation of the graph structure: in For superpixel blocks, X (1) ,X (2) ,...,X (K) The feature vector of the superpixel block;
[0020] Historical photovoltaic power data is standardized and used as independent nodes;
[0021] The edge weights are calculated using the adaptive attention matrix, as follows:
[0022] e ij=LeakyReLU(a T [W·X (i) ||W·X (j) ])
[0023]
[0024] Where W is the learnable parameter matrix, || is the connection operation, a is the attention vector, and e ij For each pair of adjacent nodes i and j, α is the attention coefficient. ij The attention coefficients after normalization;
[0025] Construct an adaptive adjacency matrix using normalized attention coefficients:
[0026] Based on ground-based cloud map segmentation, superpixel blocks, node feature vectors, and adjacency matrices are used to define nodes. and edge The set is a graph structure The formula is as follows:
[0027]
[0028] in, This is the feature vector of historical photovoltaic power values.
[0029] Furthermore, a graph sampling aggregation network is used to extract local and global coupled features in a hierarchical manner. Through mean aggregation and residual connections, dynamic graph structure features are obtained, including:
[0030] A fixed number of nodes are randomly sampled from the neighborhood, and information is extracted from the neighborhood of the target node using an aggregation function. The aggregation formula is as follows:
[0031]
[0032] in, This refers to the pooling aggregation method for adjacent nodes in the k-th layer. For the aggregated neighborhood features, N(v) represents all neighboring nodes containing the target node; W represents the updated representation of node u at layer (k-1), which serves as the input to layer k. pool b represents the learnable weight matrix during the aggregation process at the k-th layer; b represents the bias vector.
[0033] The aggregated neighborhood features are concatenated with the features from the previous layer of the target node, using the following formula:
[0034]
[0035] Among them, the CONCAT sampling neighboring nodes and their features splicing operation, Wk Let σ be the learnable weight matrix for each layer, and σ be the activation function.
[0036] The information collected through the above aggregation operations serves as the input features for subsequent prediction tasks.
[0037] Furthermore, based on photovoltaic power values, and using time-series features, image spatial features, and dynamic graph structure features as inputs, a multi-head attention mechanism is employed to fuse multimodal features and obtain a fused feature vector, including:
[0038] The graph structure feature F g Image spatial features F i Time series feature F t Each is mapped to H groups of query, key, and value vectors through linear transformation;
[0039] The cross-modal attention score is calculated using the following formula:
[0040]
[0041] Where Q, K, and V are the query, key, and value vectors of the i-th group, respectively; d k This represents the dimension of the key vector K;
[0042] The multi-attention head outputs are spliced and projected through a linear layer, as shown in the formula:
[0043]
[0044] Among them, head i =Attention(Q) i ,K i V i W is the output of the i-th attention head. O The weight matrix is the final linear transformation matrix. This is a multimodal fusion feature.
[0045] Furthermore, the time-series prediction model adopts an encoder-decoder structure. The encoder extracts deep time-series correlations, and the decoder combines multi-modal fusion features to output the predicted photovoltaic power value.
[0046]
[0047] in, For multimodal fusion features, Z t For position encoding, This is the final predicted output value of photovoltaic power.
[0048] The present invention also provides an ultra-short-term photovoltaic power prediction system based on an image-derived graph structure, comprising:
[0049] The image feature extraction module is used to acquire the ground cloud image sequence, segment the image elements using the superpixel segmentation method to obtain superpixel blocks, and extract image spatial features from the superpixel blocks using the ResNet neural network.
[0050] The dynamic graph construction module is used to construct a dynamic graph structure by using image spatial features and historical photovoltaic power time series features as nodes and adjusting edge weights based on the similarity of node features using a dynamic adjacency matrix.
[0051] The graph structure feature extraction module is used to extract local neighborhood features and global coupled features in a hierarchical manner based on dynamic graph structure using a graph sampling aggregation network. Through mean aggregation and residual connection, dynamic graph structure features are obtained.
[0052] The feature fusion module is used to perform multimodal feature fusion based on photovoltaic power values, taking time-series features, image spatial features, and dynamic graph structure features as inputs, and using a multi-head attention mechanism to obtain a fused feature vector.
[0053] The power prediction module is used to output the photovoltaic power value within a specified time interval of 5-30 minutes in the future by passing the fused feature vector through a time series prediction model.
[0054] The present invention also provides a computer device comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when the programs are executed by the processors, they implement the steps of the ultra-short-term photovoltaic power prediction method based on image-derived graph structure as described above.
[0055] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the ultra-short-term photovoltaic power prediction method based on an image-derived graph structure as described above.
[0056] Beneficial effects:
[0057] (1) This invention applies image processing using a superpixel segmentation method to aggregate adjacent pixels. This method better captures image boundaries and textures while reducing noise and computational load.
[0058] (2) This invention employs an image-derived graph construction method to convert sky images and historical data into graphs, capturing the global spatial coupling features between these patterns. This method allows for more flexible computation of coupling feature correlations, overcoming the limitations of traditional local sampling field-of-view methods.
[0059] (3) This invention applies a multimodal feature fusion method based on a multi-head attention mechanism, fusing temporal features, graphic features, and image features. This method realizes global interaction between different modal features and eliminates redundancy between sky images and historical data. Attached Figure Description
[0060] Figure 1 This is a flowchart of the photovoltaic power prediction method of the present invention;
[0061] Figure 2 This is a schematic diagram of the superpixel block acquisition process used in this invention;
[0062] Figure 3 This is a schematic diagram of the graph structure reconstruction used in this invention;
[0063] Figure 4 This is a schematic diagram of the coupled spatial feature learning module used in this invention;
[0064] Figure 5 This is a schematic diagram of the multimodal fusion module proposed in this invention;
[0065] Figure 6 This is a schematic diagram of the prediction result fitting using the proposed method in an embodiment of the present invention;
[0066] Figure 7 This chart compares the performance metrics of different prediction methods across different prediction time domains. Detailed Implementation
[0067] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0068] Reference Figure 1 The present invention provides an ultra-short-term photovoltaic power prediction method, comprising the following steps:
[0069] Step 1: Process the ground cloud image using the improved SLIC superpixel segmentation algorithm to obtain image patches with clear cloud edges and semantic coherence, and extract the depth features of the image patches using the ResNet-18 neural network.
[0070] Step 2: Construct a dynamic graph structure using image patch depth features and historical power data, where each node contains image feature vectors and power temporal features;
[0071] Step 3: Utilize the Graph-SAGE neighborhood sampling and aggregation mechanism to extract local cloud features and globally coupled features in a hierarchical manner;
[0072] Step 4: Utilize a multi-head attention mechanism to capture the complex coupling relationships between different modalities in parallel, and obtain cross-modal fusion spatiotemporal features, spatiotemporal dynamic pattern features, and graph structure features;
[0073] Step 5: Based on the fusion of multimodal features, output the photovoltaic power value for the next 5-30 minutes through a time-series prediction model.
[0074] Specifically, in this embodiment of the invention, image preprocessing is performed in step 1 using the following method.
[0075] The RGB image is converted to the CIELAB color space. A 5-dimensional feature vector [l,a,b,x,y] is calculated to represent each pixel in the image, where (l,a,b) represents the color and (x,y) represents the spatial location vector. The similarity between two pixels is measured by the distance between the 5-dimensional vectors. The distance between the vectors is Euclidean distance, which measures the direct difference between the two pixels in their CIELAB spatial locations. The formula is as follows:
[0076]
[0077] SLIC first sets a specific number of seed points, which are initially uniformly distributed on the image as cluster centers. To ensure that the seed points do not fall on noise or edge regions that may affect the segmentation accuracy, the initial seed point positions must be adjusted based on gradient optimization to avoid edge noise interference. This is done by recalculating the gradient value in the neighborhood of each seed point and repositioning it to the position with the smallest gradient in the neighborhood. The gradient calculation method is as follows:
[0078] G(x,y)=||I(x+1,y)-I(x-1,y)|| 2 +||I(x,y+1)-I(x,y-1)|| 2
[0079] Where G represents the gradient and I is the image pixel value;
[0080] By examining every pixel within the region surrounding the centroid of each superpixel block, new cluster centers are iteratively recalculated, and superpixel blocks are generated through clustering. The formula for measuring the cluster distance is:
[0081]
[0082] Where, d c For color distance, d s Where C is the spatial distance, S is the balance factor, and S is the superpixel side length.
[0083] like Figure 2 As shown, following the above operations, a ground cloud map divided into K superpixel blocks can be obtained.
[0084] For each segmented image patch, fine-grained features of the cloud image patch are extracted using a deep learning network as part of the multimodal features. According to an embodiment of the present invention, a ResNet-18 neural network is used to extract the depth features of each image patch. This is represented as: in It is a superpixel block.
[0085] The main purpose of step 2 is to construct a graph structure by combining the spatial features of image patches with historical photovoltaic power data. The edges between nodes represent the correlation between cloud cover and power time series, thereby dynamically reflecting the impact of weather changes on power output. (Refer to...) Figure 3 The features extracted from the superpixel blocks of the cloud image and the historical power data are used to construct a dynamic graph structure through an adaptive adjacency matrix. Each node contains image feature vectors and temporal power features. The specific implementation steps are as follows:
[0086] The feature vector X extracted from each superpixel block by the ResNet-18 neural network (1) ,X (2) ,...,X (K) As a node representation of a graph structure;
[0087] Historical power data was standardized and used as independent nodes.
[0088] Edge weights are calculated using an adaptive attention matrix, as shown in the formula:
[0089] e ij =LeakyReLU(a T [W·X (i) ||W·X (j) ])
[0090]
[0091] Where W is the learnable parameter matrix, || is the connection operation, a is the attention vector, and e ij For each pair of adjacent nodes i and j, α is the attention coefficient. ij The attention coefficients after normalization;
[0092] Construct an adaptive adjacency matrix using normalized attention coefficients:
[0093] Based on ground-based cloud map segmentation, superpixel blocks, node feature vectors, and adjacency matrices are used to define nodes. and edge The set is a graph structure The formula is as follows:
[0094]
[0095] in, X is the eigenvector of historical photovoltaic power values. (1) ,X (2) ,...,X (K) is the feature vector representing the node.
[0096] In step 3 of this invention, Graph-SAGE is used to extract local cloud features layer by layer through neighborhood sampling and aggregation mechanisms, ultimately capturing global spatial correlations, especially the global impact of long-distance cloud movement on photovoltaic power variations. Specifically, this includes:
[0097] like Figure 4 As shown, based on the core principle of Graph-SAGE, an aggregation function is used to sample and aggregate information from neighboring nodes, and the node embedding vector is generated by training the aggregation function.
[0098] First, a fixed number of nodes are randomly sampled from the neighborhood as the aggregation target to avoid feature imbalance caused by differences in neighborhood size. Without considering computational efficiency, all neighboring nodes can be sampled to obtain complete information.
[0099] Then, information is extracted from the neighborhood of the target node using an aggregation function. The aggregation formula is as follows:
[0100]
[0101] in, This refers to the pooling aggregation method for adjacent nodes in the k-th layer. For the aggregated neighborhood features, N(v) contains all the neighboring nodes of the target node, and u represents the neighboring nodes of the target node v. i This represents the i-th node in the neighbor set; W represents the updated representation of node u at layer (k-1), which serves as the input to layer k. pool denoted as the learnable weight matrix during the aggregation process at the k-th layer; b represents the bias vector.
[0102] After collecting neighborhood information, the aggregation function combines the embeddings of the sampled neighborhood nodes, and concatenates the aggregated neighborhood features with the features from the previous layer of the target node. The formula is as follows:
[0103]
[0104] Among them, the CONCAT sampling neighboring nodes and their features splicing operation, W k Let σ be the learnable weight matrix for each layer, and σ be the activation function.
[0105] The information collected through the above aggregation operations serves as the input features for subsequent prediction tasks.
[0106] In step 4, this invention utilizes a multi-head attention mechanism to capture the complex coupling relationships between different modalities in parallel, obtaining significant information from cross-modal fusion spatiotemporal features, spatiotemporal dynamic pattern features, and graph structure features. Specifically, such as... Figure 5 As shown, the feature dimensions from different modalities are first standardized to ensure dimensionality consistency. Then, data from all three modalities are integrated into a matrix, and attention scores are calculated using multiple attention heads, each extracting different subspace information from the fused feature matrix. The specific definitions are as follows:
[0107] The graph structure feature F g Image features F i Time series feature F t The query, key, and value vectors are mapped to H groups respectively through linear transformation, so that the multi-head cross-modal attention mechanism can uniformly align and fuse multimodal features;
[0108] The cross-modal attention score is calculated and defined as follows:
[0109]
[0110] Where Q, K, and V are the query, key, and value vectors of the i-th group, respectively; d k This represents the dimension of the key vector K, used to scale the dot product calculation result.
[0111] The multi-attention head outputs are spliced and projected through a linear layer, as shown in the formula:
[0112]
[0113] head i =Attention(Q) i ,K i V i W is the output of the i-th attention head. O The weight matrix is the final linear transformation matrix. This is a multimodal fusion feature;
[0114] In step 5 of this invention, a temporal inference model based on a progressive feature fusion architecture and a temporal convolution model is used to output the photovoltaic power prediction results for the next 10 to 30 minutes.
[0115] Step 5 specifically includes:
[0116] The fused features are processed in multiple layers through an encoder-decoder structure to further improve the accuracy of photovoltaic power prediction. In the encoder-decoder structure, the encoding part extracts deep temporal correlations, and the decoding part combines multimodal fusion features to output the predicted photovoltaic power value.
[0117]
[0118] in, For multimodal fusion features, Z t For position encoding, This is the final predicted output value of photovoltaic power.
[0119] To verify the performance of the method proposed in this invention, the following experiments were conducted in the embodiments. The dataset used was a sky imagery and PV dataset compiled by the Environmental Assessment and Optimization Group at Stanford University. The data was collected from the Green Earth Sciences Building on the Stanford University campus in the center of the San Francisco Peninsula, California (longitude 34.427°N, latitude 122.174°W). The sky image dataset includes all-sky images (128×128 resolution) taken from March 2017 to November 2019. This embodiment uses ground-based cloud images and photovoltaic power data with a time resolution of 5 minutes. After data preprocessing, approximately 30,000 samples from March 1, 2017 to March 1, 2018 were used as the training set, 3,000 samples as the validation set, and 5,000 samples randomly selected from April 1 to October 31, 2018 as the test set. Based on the above method, the ultra-short-term photovoltaic power in this region is predicted. The model performance is evaluated using quantitative error metrics. This embodiment uses four error analysis metrics: mean absolute error (MAE), normalized root mean square error (nRMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²). 2 ), defined as follows:
[0120]
[0121] Where N represents the total number of test samples, y(t) and Let these represent the predicted value and the actual value of the i-th sample, respectively. This represents the mean of the actual values.
[0122] To visually reflect the prediction results, Figure 6 The photovoltaic power generation prediction curve obtained 5 minutes in advance using the method of this invention is shown, demonstrating that the method can achieve accurate prediction of ultra-short-term photovoltaic power. To further evaluate the prediction performance of the prediction models, a total of eight comparison models were set up: LSTM, ResNet-18, 3DCNN-BiLSTM, GCN, 3D ConLSTM-CNN, transformer, PVT, and CVT. These models are all state-of-the-art ultra-short-term photovoltaic power prediction models. Furthermore, Table 1 shows the prediction error of each model for photovoltaic power prediction 5 minutes in advance. Figure 7 To compare the model constructed according to this invention with other models such as MAE, nRMSE, MAPE, and R...2 Comparison of indicators. This invention... Figure 7 The text shows what is proposed. (From Table 1 and...) Figure 7 As can be seen, compared with the comparison model, the model constructed based on this invention has smaller prediction errors and better performance.
[0123] Table 1 Comparison of prediction results from different models
[0124]
[0125] In summary, this invention realizes a novel photovoltaic power prediction method that performs multimodal learning from image-derived graphical structures, and its effectiveness has been thoroughly validated through numerical experiments. First, superpixel segmentation is used to significantly improve image preprocessing by minimizing noise interference and increasing computational efficiency. Then, an image-derived graphical construction strategy effectively converts images and historical power data into graphical structures, enabling more effective exploration of spatial coupling features through graph-based networks. Finally, a multi-head attention mechanism is used to effectively fuse multimodal data features, significantly enhancing the model's robustness and accuracy. Experimental results show that this invention performs excellently across various evaluation metrics, maintaining stable forecasting performance even under complex and variable weather conditions, which is crucial for promoting the widespread adoption of renewable energy. The experimental results demonstrate that this invention can improve the accuracy of ultra-short-term photovoltaic power prediction, providing support for the safe and stable operation of the power grid, and possesses certain engineering practical value.
[0126] Based on the same technical concept as the method embodiments, the present invention also provides an ultra-short-term photovoltaic power prediction system based on an image-derived graph structure, comprising:
[0127] The image feature extraction module is used to acquire the ground cloud image sequence, segment the image elements using the superpixel segmentation method to obtain superpixel blocks, and extract image spatial features from the superpixel blocks using the ResNet neural network.
[0128] The dynamic graph construction module is used to construct a dynamic graph structure by using image spatial features and historical photovoltaic power time series features as nodes and adjusting edge weights based on the similarity of node features using a dynamic adjacency matrix.
[0129] The graph structure feature extraction module is used to extract local neighborhood features and global coupled features in a hierarchical manner based on dynamic graph structure using a graph sampling aggregation network. Through mean aggregation and residual connection, dynamic graph structure features are obtained.
[0130] The feature fusion module is used to perform multimodal feature fusion based on photovoltaic power values, taking time-series features, image spatial features, and dynamic graph structure features as inputs, and using a multi-head attention mechanism to obtain a fused feature vector.
[0131] The power prediction module is used to output the photovoltaic power value within a specified time interval of 5-30 minutes in the future by passing the fused feature vector through a time series prediction model.
[0132] It should be understood that the ultra-short-term photovoltaic power prediction system based on image-derived graph structure in the embodiments of the present invention can realize all the technical solutions in the above method embodiments. The functions of each functional module can be specifically implemented according to the methods in the above method embodiments. The specific implementation process can be referred to the relevant descriptions in the above embodiments, which will not be repeated here.
[0133] The present invention also provides a computer device comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when the programs are executed by the processors, they implement the steps of the ultra-short-term photovoltaic power prediction method based on image-derived graph structure as described above.
[0134] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the ultra-short-term photovoltaic power prediction method based on an image-derived graph structure as described above.
[0135] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus (systems), computer devices, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0136] This invention is described with reference to a flowchart of a method according to embodiments of the invention. It should be understood that each step in the flowchart and combinations thereof can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the process. Figure 1 A device for a function specified in one or more processes.
[0137] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1The function specified in one or more processes.
[0138] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 Steps of a specified function in one or more processes.
Claims
1. A method for predicting ultra-short-term photovoltaic power based on image-derived graph structure, characterized in that, Includes the following steps: The ground cloud image sequence is obtained, and the image features are segmented using the superpixel segmentation method to obtain superpixel blocks. The image spatial features are then extracted from the superpixel blocks using the ResNet neural network. Using image spatial features and historical photovoltaic power time series features as nodes, a dynamic graph structure is constructed by adjusting edge weights based on the similarity of node features using a dynamic adjacency matrix. Based on the dynamic graph structure, a graph sampling aggregation network is used to extract local neighborhood features and global coupled features in a hierarchical manner. Through mean aggregation and residual connection, dynamic graph structure features are obtained. Based on photovoltaic power values, and using time-series features, image spatial features, and dynamic graph structure features as inputs, a multi-head attention mechanism is used to fuse multi-modal features to obtain a fused feature vector. The fused feature vectors are used to output photovoltaic power values for a specified time interval within the next 5-30 minutes through a time-series prediction model. Specifically, a graph sampling aggregation network is used to extract local and global coupled features in a hierarchical manner. Through mean aggregation and residual connections, dynamic graph structure features are obtained, including: A fixed number of nodes are randomly sampled from the neighborhood, and information is extracted from the neighborhood of the target node using an aggregation function. The aggregation formula is as follows: in, This refers to the pooling aggregation method for adjacent nodes in the k-th layer. For the neighborhood features of aggregation, This includes all neighboring nodes of the target node; This represents the updated representation of node u at layer (k-1), which serves as the input to layer k. b represents the learnable weight matrix during the aggregation process at the k-th layer; b represents the bias vector. The aggregated neighborhood features are concatenated with the features from the previous layer of the target node, using the following formula: in, The operation of concatenating sampling neighborhood nodes and their features. For each layer, there is a learnable weight matrix. For activation functions; The information collected through the above aggregation operations serves as the input features for subsequent prediction tasks.
2. The method according to claim 1, characterized in that, Image feature segmentation is performed using superpixel segmentation methods to obtain superpixel blocks, including: Convert the RGB image to the CIELAB color space and calculate the 5-dimensional feature vector. To represent each pixel in the image, where Indicates color, Representing spatial location vectors, the similarity between two pixels is measured by the distance between the 5-dimensional vectors; The initial seed point position is adjusted based on gradient optimization, and superpixel blocks are generated by iterative clustering.
3. The method according to claim 2, characterized in that, The distance metric formula for clustering is: in, For color distance, For spatial distance, As a balance factor, Let be the superpixel side length, and D be the distance between cluster centers.
4. The method according to claim 1, characterized in that, Using image spatial features and historical photovoltaic power time-series features as nodes, a dynamic graph structure is constructed by adjusting edge weights based on the similarity of node features using a dynamic adjacency matrix, including: The feature vector of each superpixel block extracted by the ResNet neural network is used as a node representation of the graph structure: ,in For superpixel blocks, The feature vector of the superpixel block; Historical photovoltaic power data is standardized and used as independent nodes; The edge weights are calculated using the adaptive attention matrix, as follows: in, For learnable parameter matrix, For connection operations, For attention vectors, For each pair of adjacent nodes i and j, the attention coefficients are... The attention coefficients after normalization; Construct an adaptive adjacency matrix using normalized attention coefficients: ; Based on ground-based cloud map segmentation, superpixel blocks, node feature vectors, and adjacency matrices are used to define nodes. and edge The set is a graph structure The formula is as follows: in, This is the feature vector of historical photovoltaic power values.
5. The method according to claim 1, characterized in that, Based on photovoltaic power values, and using time-series features, image spatial features, and dynamic graph structure features as inputs, a multi-head attention mechanism is employed to fuse multimodal features and obtain a fused feature vector, including: graph structure features Image spatial features Time sequence features Mapped to respectively through linear transformation Grouped queries, key-value vectors; The cross-modal attention score is calculated using the following formula: in, These are the query, key, and value vectors for the i-th group, respectively. Represents the key vector The dimension; The multi-attention head outputs are spliced and projected through a linear layer, as shown in the formula: in, For the output of the i-th attention head, The weight matrix is the final linear transformation matrix. This is a multimodal fusion feature.
6. The method according to claim 1, characterized in that, The time-series prediction model uses an encoder-decoder structure. The encoder extracts deep time-series correlations, while the decoder combines multi-modal fusion features to output the predicted photovoltaic power value. in, For multimodal fusion features, For position encoding, This is the final predicted output value of photovoltaic power.
7. An ultra-short-term photovoltaic power prediction system based on image-derived graph structure, characterized in that, include: The image feature extraction module is used to acquire the ground cloud image sequence, segment the image elements using the superpixel segmentation method to obtain superpixel blocks, and extract image spatial features from the superpixel blocks using the ResNet neural network. The dynamic graph construction module is used to construct a dynamic graph structure by using image spatial features and historical photovoltaic power time series features as nodes and adjusting edge weights based on the similarity of node features using a dynamic adjacency matrix. The graph structure feature extraction module is used to extract local neighborhood features and global coupled features in a hierarchical manner based on dynamic graph structure using a graph sampling aggregation network. Through mean aggregation and residual connection, dynamic graph structure features are obtained. The feature fusion module is used to perform multimodal feature fusion based on photovoltaic power values, taking time-series features, image spatial features, and dynamic graph structure features as inputs, and using a multi-head attention mechanism to obtain a fused feature vector. The power prediction module is used to output the photovoltaic power value within a specified time interval of 5-30 minutes in the future by passing the fused feature vector through a time series prediction model. Specifically, a graph sampling aggregation network is used to extract local and global coupled features in a hierarchical manner. Through mean aggregation and residual connections, dynamic graph structure features are obtained, including: A fixed number of nodes are randomly sampled from the neighborhood, and information is extracted from the neighborhood of the target node using an aggregation function. The aggregation formula is as follows: in, This refers to the pooling aggregation method for adjacent nodes in the k-th layer. For the neighborhood features of aggregation, This includes all neighboring nodes of the target node; This represents the updated representation of node u at layer (k-1), which serves as the input to layer k. b represents the learnable weight matrix during the aggregation process at the k-th layer; b represents the bias vector. The aggregated neighborhood features are concatenated with the features from the previous layer of the target node, using the following formula: in, The operation of concatenating sampling neighborhood nodes and their features. For each layer, there is a learnable weight matrix. For activation functions; The information collected through the above aggregation operations serves as the input features for subsequent prediction tasks.
8. A computer device, characterized in that, include: One or more processors; Memory; And one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when the programs are executed by the processors, they implement the steps of the ultra-short-term photovoltaic power prediction method based on image-derived graph structure as described in any one of claims 1-6.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the ultra-short-term photovoltaic power prediction method based on image-derived graph structure as described in any one of claims 1-6.